A Weakly Supervised Soil Pore Segmentation Method Based on Traditional Segmentation Algorithm

Soil pore structure plays an important role in the ecosystem. In recent years, researchers have begun utilizing deep learning to segment soil pores. However, when confronted with a large number of soil pore datasets that require annotation, the effort and time for manual labeling are limited and insufficient to accurately annotate the entire dataset. To address this issue, this paper proposes a weakly supervised soil pore segmentation method (WSSPS) based on traditional segmentation algorithms. WSSPS generates soil pore pseudo-labels through the traditional segmentation algorithm for pre-training in the upstream task. Subsequently, fine-tuning was performed in the downstream task using expert-defined labels that only accounted for 1.8% to 35.6% of the total dataset to obtain the final segmentation effect map. In this study, three traditional segmentation algorithms are utilized for comparison experiments in the upstream task, and they are also compared with each other and four supervised deep learning methods. The results demonstrate that WSSPS not only possesses better segmentation results than traditional and supervised methods, but also greatly reduces the amount of manual annotation. This study facilitates the application of deep learning in soil pore segmentation and provides image processing technical support for the advancement of modern soil research.